Spatial-guided frequency cross-attention with gated single-branch decoding for 3D brain tumor segmentation

Abstract

Automatic segmentation of glioma subregions in multimodal magnetic resonance imaging remains challenging because enhancing tumor and tumor core regions are often small, heterogeneous, and weakly separated from surrounding tissues. This paper presents an improved frequency-aware 3D segmentation network built upon HFF-Net. The model keeps the low-frequency and high-frequency dual-encoder design of HFF-Net, while replacing the original frequency-domain crossattention with a spatial-guided frequency cross-attention module. Different from the original frequency-domain cross-attention, which mainly performs branch interaction in the Fourier domain without an explicit tumor-location constraint, the proposed module first learns a multi-semantic 3D spatial prior and uses it to make frequency modulation tumor-aware. The prior modulates both real and imaginary Fourier components, after which frequency-domain enhancement, progressive channel self-attention, and cross-slice consistency modeling are applied in a structured cascade. Clinically, low-frequency cues correspond to global tumor extent, edema distribution, and smooth anatomical context, whereas high-frequency cues correspond to tumor boundary details, enhancement rims, and heterogeneous internal structures. In addition, the decoder is redesigned as a gated single-branch decoder: features from low- and high-frequency branches are adaptively merged at multiple scales before being decoded through one unified reconstruction path. This design preserves complementary frequency cues while reducing the ambiguity caused by separate decoder outputs. The experimental protocol follows the HFF-Net setting on BraTS2019 and BraTS2020.